Overview
- Tool name: DeepSeek V4 (deepseek-v4-flash / deepseek-v4-pro)
- Developer: DeepSeek
- Official website: https://api-docs.deepseek.com
- Category: AI Models
- License: MIT (open weights)
DeepSeek V4 is the current two-tier lineup from DeepSeek, and among the most aggressively priced frontier-class APIs on the market:
- DeepSeek V4-Flash: the workhorse for general chat, agents and tool calling; supports thinking and non-thinking modes (thinking by default), with a 1M-token context and up to 384K output
- DeepSeek V4-Pro: the reasoning and complex-code tier; also 1M context, built for math, code and long-horizon tasks in thinking mode
Both tiers support JSON output, tool calls, chat prefix completion and FIM completion, and offer OpenAI- and Anthropic-compatible endpoints (api.deepseek.com and api.deepseek.com/anthropic), making migration nearly free. V4 weights are released under MIT for unrestricted commercial use and self-hosting. The legacy aliases deepseek-chat / deepseek-reasoner were retired on July 24, 2026, mapping to V4-Flash's non-thinking / thinking modes.
API pricing (official, per million tokens)
- V4-Flash: $0.14 input ($0.0028 cache hit) / $0.28 output
- V4-Pro: $0.435 input ($0.003625 cache hit) / $0.87 output
- Concurrency: 2,500 for Flash / 500 for Pro
- DeepSeek has announced a peak/off-peak pricing policy (peak at 2x regular price; peak hours are 9-12 and 14-18 Beijing time on weekdays) - see the official announcement for the effective date
Pricing verified against DeepSeek's official docs at api-docs.deepseek.com (October 2026).
Who it is for
- High-throughput production: input costs roughly 1/20 of GPT-5.6 at comparable capability
- Agents and tool calling: 2,500 concurrency plus near-free cache hits suits multi-turn agent loops
- Very long context: 1M input with 384K output for whole codebases and long-document analysis
- Self-hosting: MIT weights with full vLLM/SGLang ecosystem support
Limitations
- Multimodal (image) input remains experimental (V4-Flash-Vision-Exp)
- Once peak/off-peak pricing lands, peak-hour costs double; real-time businesses should model this
- Pro's 500-concurrency limit means large batch jobs may queue or need the Batch API
Frequently asked questions
What does DeepSeek V4 cost?
Official pricing: Flash at $0.14 input / $0.28 output; Pro at $0.435 input / $0.87 output (per million tokens), with cache-hit input as low as $0.0028. Note the announced peak/off-peak policy (2x at peak).
Is DeepSeek V4 open source?
Yes. Weights are MIT-licensed for commercial use, fine-tuning and self-hosting with no strings attached.
Is deepseek-chat still available?
It was retired on July 24, 2026. Migration: deepseek-chat to deepseek-v4-flash (non-thinking), deepseek-reasoner to deepseek-v4-flash (thinking).
DeepSeek V4 vs Kimi K3 vs GLM-5?
DeepSeek V4 for the lowest cost and highest concurrency; Kimi K3 for a 1M-context open-weights coding flagship; GLM-5 for a balanced MIT flagship trained on domestic Chinese hardware.
*This page is not affiliated with DeepSeek. Pricing verified as of October 2026 (api-docs.deepseek.com); check official announcements for the peak/off-peak effective date.*